Is 2025 the Year AI Takes Over? Inside Nvidia’s Dual‑Platform Revolution

Dual‑Platform Shift in Computing

At CES 2026 Jensen Huang described a historic dual‑platform transformation where accelerated computing and generative AI will explode simultaneously, not merely as technology upgrades but as a full‑stack reconstruction expected to unlock $10 billion of modern‑computing value.

From Code on CPUs to Models on GPUs

Huang argued that traditional programming—writing code and running pre‑compiled binaries on CPUs—is ending. Future workloads will train models and generate each pixel and token in real time on GPUs.

AI‑First Development

The industry is moving from application‑first to AI‑first, reallocating R&D budgets from conventional methods to AI‑centric approaches.

2025: AI’s “Ascension Year”

Test‑Time Scaling : real‑time reasoning capabilities.

Agent Systems : inference, planning, and tool‑use abilities.

Physical AI : understanding physics and interacting with the real world.

Open‑Source Ascendant : simultaneous innovation across companies and industries.

Open‑Source Model Push

Nvidia is building its own AI supercomputers and has released several heavyweight open‑source models:

Nemo Tron 3 : language model using a mixed SSM architecture for ultra‑fast inference.

OpenFold 3 : biology model that understands and generates protein structures.

ForecastNet : a new approach to weather prediction.

Huang emphasized the mantra “model + data = trust,” stressing that both model capability and data quality are essential for user confidence.

AI Agents as the New Atomic Unit

Future AI will evolve from memory‑only systems to AI agents capable of reasoning. The breakthrough is Agent Routing , where a single model is insufficient and a mixture‑of‑experts approach is required. Nvidia provides a blueprint using the NVIDIA NEMO framework, an Agent Router, and APIs to create fully customizable, cutting‑edge AI solutions.

An illustrative example is the personal AI assistant “RICCI,” built on a local Dell DGX Spark, cloud APIs (including ElevenLabs for voice), and open‑source models. RICCI can manage schedules, generate full‑rendered content from sketches, and even control robots, showcasing AI’s transition from digital to physical domains.

Physical AI Revolution

Current AI lacks physical commonsense—causality, friction, inertia, gravity—knowledge that even toddlers possess. Huang proposed a three‑tier Physical AI architecture: training computers, inference computers, and simulation computers. The simulation tier is foundational, enabling perception of physical feedback.

Nvidia’s COSMOS platform delivers generative physical‑AI capabilities, aligning language, images, 3D, and motion to produce photorealistic, physically coherent videos, and supports interactive closed‑loop simulation, inference, and prediction for rapid robot adaptation.

Partnering with Siemens, Nvidia integrates the full AI stack across design, production, and operations, turning factories into giant robots.

VERA RUBIN Platform for the Trillion‑Parameter Era

Facing a ten‑fold annual growth in AI model size and a five‑fold increase in token generation, traditional architectures are obsolete. Nvidia’s VERA RUBIN platform introduces:

RUBIN GPU : 1.7× more transistors than Blackwell, with MVF P4 tensor engine delivering 1.6× performance boost.

VERA CPU : double the performance‑per‑watt of Grace, 88 cores, 176 threads, and high I/O.

NVLink 6 & Spectrum‑X Ethernet : 240 TB/s intra‑rack bandwidth at near‑zero cost.

Bluefield‑4 DPU : standard on every node, offloading virtualization, security, and networking tasks.

Overall performance jumps include 5× peak inference and 3.5× training speed. To overcome KV‑cache memory and bandwidth bottlenecks, VERA RUBIN adds silicon‑photonic links and distributed context memory, scaling GPU context to 16 TB and reshaping data‑center architecture.

Energy efficiency and security also improve dramatically: double the performance‑per‑watt, constant‑temperature operation at 45 °C, full‑link encryption, and 100 % power‑budget utilization. Training a 10‑trillion‑parameter model now requires only a quarter of the systems needed for Blackwell, and factory throughput is 100× higher than Hopper.

Conclusion

Huang concluded that Nvidia has become the world’s leading AI engine, driving a new industrial revolution from chips to systems, cloud to robots, and digital to physical worlds. The “alchemy of computing” is underway, and we stand at a historic turning point.